Exploratory Hierarchical Clustering for Management Zone Delineation in Precision Agriculture

Transcription

1 Exploratory Hierarchical Clustering for Management Zone Delineation in Precision Agriculture Georg Ruß, Rudolf Kruse Otto-von-Guericke-Universität Magdeburg, Germany ICDM, Sep 2nd, 2011

2 Precision Agriculture GPS technology used in site-specific, sensor-based crop management combination of agriculture and information technology data-driven approach to agriculture lots of data analysis tasks

3 Data Details Example Field Figure: F550 field, depicted on satellite imagery, source: Google Earth

4 Data Details Features collect a number of geo-coded, high-resolution features such as: N1, N2, N3: nitrogen fertilizer application rates in 2004 REIP32, REIP49: vegetation index (red edge inflection point) in 2004 Yield: corn yield 2003, winter wheat yield in 2004 and 2007 EC25: electrical conductivity of soil in 2004 ph, P, K, Mg: soil sampling in 2007 one field available, 1080 records in 25 25m-resolution on a hexagonal grid

5 Data Details Temporal Aspects time EC25 Yield 2003 N1 Yield 2004 REIP49 / N3 REIP32 / N2 Yield 2007 ph / P / K / Mg Figure: timeline of data acquisition

6 Spatial Autocorrelation Are (spatial) data records independent of each other? (Do we have spatial autocorrelation?) [17.36,27.43] (27.43,30.06] (30.06,31.77] (31.77,32.64] (32.64,34.37] (34.37,35.84] (35.84,37.36] (37.36,38.94] (38.94,40.89] (40.89,80.5] (a) EC25 [4.8,8.7] (8.7,9.2] (9.2,9.5] (9.5,9.8] (9.8,10] (10,10.3] (10.3,10.6] (10.6,11] (11,11.6] (11.6,18.3] (b) Magnesium content Figure: F550, EC25 and Magnesium readings

7 Management Zone Delineation A common task in agriculture: subdivide the field into smaller zones zones are rather homogeneous zones are spatially mostly contiguous similarity between zones is low spatial clustering

8 Literature Approaches mostly non-spatial algorithms are used no spatial contiguity small islands, outliers, etc. black-box models fuzzy c-means, k-means, etc. spatial contiguity is not always required, but desirable spatial autocorrelation is usually neglected rather than exploited

9 Spatial Contiguity Constraint spatial clustering = clustering with a spatial contiguity constraint constrained clustering Keep it simple and understandable: hierarchical clustering agglomerative clustering Idea: 1. split field into small zones which are homogeneous 2. iteratively merge these zones obeying similarity and spatial constraint

10 Spatial Tessellation k-means clustering on the data points coordinates due to spatial autocorrelation, adjacent points are likely to be similar this ensures homogeneity of these small zones k is user-controllable and easy to understand homogeneous field: smaller k heterogeneous field: higher k

11 Spatial Tessellation F550, 80 zones, EC25 Northing Easting [17.36,23.59] Figure: Tessellation of F550 using k-means, k = 80 (grey shades are for illustration only, no further meaning here)

12 Hierarchical Agglomerative Constrained Clustering principle: merge only adjacent objects/clusters, if they are similar enough this ensures spatial contiguity spatial constraint, non-adjacent clusters cannot link once non-adjacent clusters become much more similar than adjacent ones, they may be merged introduce a user-controllable contiguity factor cf cf 2: high contiguity cf [1,2]: low contiguity cf 1: no contiguity

13 HACC 1D example [17.36,27.43] (27.43,30.06] (30.06,31.77] (31.77,32.64] (32.64,34.37] (34.37,35.84] (35.84,37.36] (37.36,38.94] (38.94,40.89] (40.89,80.5] (a) EC25 plot (b) EC25, 80 zones (c) EC25, final zones Figure: F550, EC25 clustering

14 HACC 4D example [4.5,5.4] (5.4,5.5] (5.5,5.6] (5.6,5.7] (5.7,5.8] (5.8,6] (6,6.2] (6.2,6.4] (6.4,6.9] (6.9,8.9] (a) ph plot [1,4.3] (4.3,4.8] (4.8,5.3] (5.3,6.1] (6.1,6.7] (6.7,7.6] (7.6,9] (9,11] (11,14.9] (14.9,58.8] (b) P plot [4.8,8.7] (8.7,9.2] (9.2,9.5] (9.5,9.8] (9.8,10] (10,10.3] (10.3,10.6] (10.6,11] (11,11.6] (11.6,18.3] (c) Mg plot [5.4,7.7] (7.7,8.6] (8.6,9.3] (9.3,10.1] (10.1,11.2] (11.2,12.3] (12.3,14.1] (14.1,16.3] (16.3,22.5] (22.5,73.7] (d) K plot Figure: F550, four attributes

15 HACC 4D example (cont.) (a) F550-4D, beginning (b) F550-4D, ten zones Figure: F550, management zones actually, 3 zones (when comparing attribute values) low ph, low P, low Mg, low K (largest zone) high ph, high P, high Mg, high K (border zones) high ph, high P, low Mg, high K (middle, from left)

16 Summary precision agriculture as a data-driven approach spatial, geo-referenced data records in large amounts management zone delineation solved as a spatial clustering approach important difference between spatial and non-spatial data treatment use models which are fit for spatial tasks

17 Time for... Questions? Workshop Data Mining in Agriculture on Saturday (after ICDM) contact: slides, R scripts and further info at